Platonic Representations for Poverty Mapping: Unified Vision-Language Codes or Agent-Induced Novelty?

Summary

This paper talks about a new method that uses both satellite images and text data together to better predict how wealthy or poor households are, and it shows that using text generated by large language models works better than text collected by agents.

What's the problem?

The problem is that models that only use satellite images to estimate poverty aren’t always very accurate, and it’s hard to add different kinds of information like text in a way that improves predictions.

What's the solution?

The paper introduces a framework that combines visual data from satellites with language information, using codes that represent both types together, and finds that text created by language models is more helpful than text gathered by retrieval systems for predicting poverty.

Why it matters?

This matters because better poverty mapping helps governments and organizations target aid and resources more effectively to people who need it most, improving the impact of social programs and economic planning.